ResNet-Based 15-Arcsec Bathymetric Prediction in the Northwest Pacific Using SWOT-Derived Gravity Data and Multibeam Observations

To fully exploit the constraining capability of high‑accuracy multibeam echo sounding (MBES) data for bathymetric modeling, this study proposes a deep residual network (ResNet)-based framework for direct bathymetric prediction on a 15-arcsec grid. The framework integrates multisource gravity features derived from the Surface Water and Ocean Topography (SWOT) mission, including gravity anomaly, vertical gravity gradient, and deflection of the vertical, while using MBES data as supervisory information to construct bathymetric models for local regions in the northwestern Pacific Ocean. A ResNet architecture with three stages and six residual blocks is adopted to learn the nonlinear relationship between gravity-field information and bathymetry at wavelengths shorter than 160 km, while the SYSU_Topo model is incorporated to provide the long-wavelength background component. Results show that the proposed framework better resolves fine-scale bathymetric features and fills data-void regions, substantially mitigating the blurring effect in areas lacking shipborne survey tracks. Evaluation on the independent test set shows that the ResNet model achieves an RMSE of 70.13 m and a relative accuracy of 0.83% in Area B (163°E–169°E, 16°N–22°N). In Area C (130°E–140°E, 5°N–15°N), it improves accuracy by approximately 17% relative to GEBCO_2025 and topo_27.1, and outperforms SDUST2023BCO, BATHY-FE, and SYSU_Topo. Cross-region tests show more stable performance for the selected ResNet–feature combination, while Gaussian distance-weight encoding reduces transfer errors and systematic biases relative to absolute coordinates. Overall, the proposed method improves the short- and medium-wavelength bathymetric representation of SYSU_Topo and provides a promising pathway for refining global seafloor topography.